Atmospheric distillation column is one of the most important units in an oil refinery where crude oil that is little use as it is, is fractioned into its more valuable constituents. Different fractions of the crude distillation unit are then further processed in downstream conversion and purification units to produce final products such as gasoline, diesel and jet fuel. The physical properties and the quantity of the final products vary depending on the physicochemical properties of the crude oil being processed and the operation parameters of the atmospheric distillation column. The process operators should keep the physical properties of hydrocarbon products in specified limits and operate the crude distillation unit according to instructions from the production planning department for maximizing the profit from operations. The physical properties of different crude distillation unit fractions must be measured by taking a sample from the stream periodically and analyzing these samples in a laboratory with appropriate equipment or by online analyzers that are very expensive to install, operate and maintain. Almost all of the state-of- the art online equipment has a time lag to complete the analysis in real time due to complexity of the analyses. Therefore, the laboratory or on-line measurements become available to decision makers and operator with a time lag. The intermittent nature of the measurements leads to sub-optimal control of the unit or in some cases leads to off-spec products that have considerably less economic value than the products that are within specified limits. As a result, estimation of the physical properties from online plant data and implementation of a soft sensor has a potential benefit in improving the profit. Among different approaches, the support vector regression shows great promise in machine learning due to its ability in generalizing well to unseen test data.
We aim to estimate the physical properties of the hydrocarbon products of an atmospheric distillation column using the large volume of data obtained from real-time and online sensors installed in the system. Linear, Polynomial and Gaussian Radial Basis Function (Gaussian RBF) kernels have been tested and the SVR parameters have been optimized by embedding k-fold cross validation using a variety of algorithms including genetic algorithm (GA), grid search (GS) and non-linear programming (NLP). The performance of SVR has been compared against artificial neural network (ANN) method and robust quality estimator (RQE) already functioning in the ADC. The testing results suggest that SVR with any of the integrated kernels performs well in estimating the property and generalizes well to blind test data. Compared to RQE, the mean testing error of estimation has been improved by 31% with SVR from 6.4ËšC to 4.4ËšC and the standard deviation of estimation error is improved by 23% with SVR from 7.3 ËšC to 5.6 ËšC.
